All aboard: the effects of port development

HAL Id: halshs-03041845

https://halshs.archives-ouvertes.fr/halshs-03041845

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

All aboard: the effects of port development

César Ducruet, Juhasz Reka, David Krisztian Nagy, Claudia Steinwender

To cite this version:

César Ducruet, Juhasz Reka, David Krisztian Nagy, Claudia Steinwender. All aboard: the effects of port development. 2020. �halshs-03041845�

All aboard: The effects of port development

∗

César Ducruet

Réka Juhász

Dávid Krisztián Nagy

Claudia Steinwender

March, 2021

Abstract

By using local land intensively, ports put pressure on land prices and crowd out other economic activity. Using the introduction of containerized shipping – a relatively land-intensive technology – we find an important role for this effect. At the local level, we find that the causal effect of the shipping boom caused by containerization on pop-ulation is zero – port development increases city poppop-ulation by making a city more attractive, but this well-known market access effect is fully offset by the crowding-out mechanism. At the aggregate level, while we find overall welfare gains from container-ization, our quantitative model featuring endogenous port development also implies i) sizeable welfare costs associated with the increased land-usage of ports, and ii) size-able gains from cities’ endogenous specialization across port- and non-port activities. These mechanisms are particularly important for targeted port development policies, which we illustrate using the Maritime Silk Road.

JEL: R40, O33, F6

Keywords: Port development, Containerization, Quantitative Economic Geography

∗_{We thank Treb Allen, David Atkin, Leah Brooks, Arnaud Costinot, Don Davis, Dave} Don-aldson, Joseph Doyle, Nicolas Gendron-Carrier, Matt Grant, Gordon Hanson, Tarek Hassan, Tom Holmes, David Hummels, Amit Khandelwal, Giampaolo Lecce, Nels Lind, Nina Pavc-nik, Giacomo Ponzetto, Jim Rauch, Steve Redding, Roberto Rigobon, Andrés Rodríguez-Clare, Esteban Rossi-Hansberg, Daniel Sturm, Tavneet Suri, Jaume Ventura, Jon Vogel and David Weinstein for helpful comments and discussions. We thank Bruce Blonigen and Mario Martín Antón and their co-authors for kindly sharing data. We thank staff at the Port Authorities of Houston, New Orleans, Long Beach, Los Angeles, Portland, San Francisco and Seattle for help-ing with our information requests. Olalekan Bello, Sabrina Chen, Naman Garg, Yi Jie Gwee, Hamza Husain, Felix Iglhaut, Rodrigo Martínez Mazza, Emanuela Migliaccio, Shuhua Si, Yue Yu, Howard Zihao Zhang and a team of Columbia University undergraduate students provided outstanding research assistance. César gratefully acknowledges financial support from the ERC (Starting Grant ’World Seastems’ No. 313847). Réka is grateful for financial support from the Provost’s Office at Columbia.

Introduction

From Sri Lanka to the Netherlands, countries across the income distribution invest

heav-ily in port development.1 _{Seaports play a vital role in the global trading system,}

han-dling over 80% of world merchandise trade in 2018 in terms of volume (UNCTAD,

2019). Rich and poor countries alike view investments into ports as an integral part of

their growth strategy, as modern facilities allowing for the fast flow of cargo through the

port are a precondition for a country to participate in global production networks (

Ro-drigue,2016, p. 131). Despite this, ports have been understudied relative to other forms

of transport infrastructure such as roads or railways.2 _{In particular, little is known about}

the economic effects of port development. What determines the economic geography of ports (i.e., where port activity is located)? What are the gains from port development and how are they distributed across space?

In this paper, we study these questions by examining a major technological shock to port development: the introduction of containerization (the handling of cargo in stan-dardized boxes). We find that cities exogenously more suited to this new technology witnessed a boom in shipping flows after the onset of containerization, but not before. Surprisingly, however, this boom in local shipping did not translate into population inflows: we find an effect of shipping on population in our IV estimates that is both economically and statistically insignificant. To conduct this analysis, we use a unique dataset of city populations and shipping flows worldwide for the period 1950-1990 to estimate the local, city-level effects of containerization. To isolate exogenous variation, we build on a previous literature that has shown that access to deep sea ports was an

important determinant of a city’s suitability for containerization (Brooks et al., 2019;

Altomonte, Colantone, and Bonacorsi, 2018). We develop a novel measure of ‘natu-rally endowed’ depth (as distinct from depth attained by dredging) using granular data on oceanic depths around each city in our data.

We view the zero local population effects of containerization as an unexpected find-ing. It is in contrast to standard models that predict an inflow of population as improved

market access makes a location more desirable for firms and consumers (Co¸sar and

Fa-1_{For example, the Port of Rotterdam (Netherlands) undertook the expansion of its container} facilities by 110 ha in 2004 at a cost of EUR 657m, 200m of which was financed by the European Investment Bank (Source:https://www.eib.org/en/projects/pipelines/all/20030288). The Port of Colombo (Sri Lanka) has made massive investments in recent years. A single project upgrading harbor infrastructure was undertaken between 2008-2012 at a cost of Rs 42 billion (Source:

https://www.slpa.lk/port-colombo/projects).

2_{Redding and Turner(2015) provide an overview of this literature. An exception isBrooks,}

Gendron-Carrier, and Rua(2019), who study the reduced-form effects of containerization on county-level economic outcomes in the U.S.

jgelbaum, 2016;Nagy, 2020b;Fajgelbaum and Redding, 2018). Indeed, other papers studying similar shocks to a location’s accessibility have found a positive effect on

pop-ulation (Bleakley and Lin,2012;Campante and Yanagizawa-Drott,2018;Brooks et al.,

2019).

We argue that the zero local population effects are driven by a crowding-out mech-anism. Different to other transport infrastructures such as railways or roads, ports are investments that occupy large amounts of land in the cities in which they are located. Today, the median port in our worldwide sample occupies space equivalent to 250

soc-cer pitches, while the port at the top decile occupies 1,100.3 By using locally scarce

land resources heavily, ports may drive up land rents and crowd out other economic activity.

Using rich historical evidence, as well as systematic contemporary data on port area and cargo composition, we show that containerization is indeed a much more land-intensive technology than the one it replaced. Our analysis suggests that moving from a fully non-containerized port to a fully containerized one requires 75% larger land area (holding the volume of traffic constant). As such, the higher land intensity of containerized port technology can provide an explanation for the zero population effect. Intuitively, the increased use of scarce local land can counteract the market access effect by driving up land prices and crowding out other economic activity from the city. We provide empirical evidence for this mechanism by showing that after containerization, shipping increased disproportionately more in low land rent cities.

Of course, the land-price mechanism is not the only force that can lead to the crowd-ing out of population. We consider three other prominent mechanisms that could ac-count for our findings; i) the lower labor intensity of containerized port technology, ii) pollution and other disamenities associated with port development, and iii) additional decreases in overland transportation costs caused by containerization. We show that these either do not stand up to more rigorous examination, or are quantitatively too small to explain our results.

Informed by the local, reduced form effects of port development, in the second part of the paper we develop a model to study the effects of port development in general equilibrium. The model is an otherwise standard economic geography model of trading cities to which we add an endogenous port development decision. As such, the model incorporates not only the standard market access effect, but also allows for port devel-opment to crowd out other forms of economic activity. This is because in the model,

3_{We use high resolution remote sensing data from Google Earth to delineate the area} occu-pied by a random subset of 236 ports in our dataset. We discuss the methodology in detail in the Supplementary Material.

developing the port (and hence reducing trade costs) requires scarce local land that can be used for other purposes. Whether a city ultimately gains in population is the out-come of the trade-off between the market access and crowding-out mechanisms. Thus, the model has the ability to rationalize the zero population effects of shipping found in the data.

Guided by the model, we re-estimate the causal effect of increased shipping flows on population controlling for market access. In line with the model predictions, our causal estimates point to a negative effect of shipping on city population once market access is controlled for. This finding provides further empirical evidence consistent with the crowding-out effect of port development.

Next, we examine the aggregate effects of port development by taking the model to the data. We use data on shipping flows, city GDP and population in 1990 to back out cities’ unobserved model fundamentals. We conduct two counterfactual simulations. In our first counterfactual, we simulate the pre-containerization equilibrium in the model by undoing the containerization shock. We show that the model-simulated data closely match the zero population effects of shipping using the same IV strategy (based on depth) as in the reduced form. Furthermore, we show that containerization increased shipping more in low land-rent cities, as in the data.

Our estimates suggest that containerization increased world welfare by 3.84%. To better understand how the crowding-out channel affects these welfare gains, we com-pare the aggregate welfare effects in our model to what a standard model in which transport cost reductions are exogenous and free (i.e., they do not use scarce resources) would predict. We find a quantitatively meaningful role for two mechanisms. First, we estimate the aggregate resource cost of containerization to be substantial: it reduces the welfare gains arising from a standard model by about 16%. Second, we also find a role for additional welfare gains stemming from endogenous specialization in port-and non-port activities based on comparative advantage. In particular, these gains offset about 58% of the resource cost of containerization. In addition, we find that, unlike in our baseline model, the local population effects of shipping are positive, economically meaningful and statistically significant in the standard model. This result again under-scores the link between the zero local population effects of shipping and the endogenous crowding-out mechanism that is absent from standard models.

In our second counterfactual, we examine the effects of targeted port-development policies. We focus on a setting similar to the ‘Maritime Silk Road’ project – a large set of port investments currently being undertaken by China in South-Asian, African and European ports. Our findings suggest that targeted port development has the po-tential for large distributional effects triggered by the reallocation of shipping activity.

Most strikingly, we predict a large decline in shipping in Singapore (a non-targeted port which we estimate loses about 50% of its shipping flows), which is driven by the fact that shipping activity reallocates to nearby, targeted ports. The initial shock is then am-plified by less endogenous port development in Singapore as demand for port services falls, illustrating the increasing returns to scale at work in our model. However, despite losing a sizeable fraction of its shipping flows, Singapore gains 1% in GDP, as resources reallocate to Singapore’s highly productive non-port activities. This illustrates that, be-cause of the resource cost of port development, gains and losses in shipping do not translate directly into gains to real GDP. These findings highlight the importance of ac-counting for our endogenous port development mechanism when quantifying how the gains from targeted port development are distributed across space. More speculatively, they question the wisdom of highly productive, expensive cities such as Hong Kong and Singapore continuing to specialize heavily in port services.

Related literature. A recent, growing literature provides evidence that better trading

opportunities lead to local benefits inducing city development (Bleakley and Lin,2012;

Armenter, Koren, and Nagy, 2014; Nagy, 2020a; Campante and Yanagizawa-Drott,

2018). Some of these studies focus on city development at port locations in

partic-ular (Fujita and Mori, 1996; Co¸sar and Fajgelbaum, 2016; Fajgelbaum and Redding,

2018). We contribute to this literature by showing that trade-induced development can

also have substantial local costs. The crowding-out mechanism that drives the cost side in our setting also relates the paper to the ‘Dutch disease’ literature. This literature shows that booming industries can entail significant costs by putting a strain on scarce

local resources and therefore crowding out other (tradable) sectors (Corden and Neary,

1982;Krugman,1987;Allcott and Keniston,2017).4 Relative to this literature, our set-ting contains the potential for not only costs but also gains, as booming port activities benefit local tradables through improving market access. Thus, one contribution of our paper is to generalize the predictions from these two, seemingly disparate literatures that have focused on either the costs or the benefits from booming sectors.

Our paper is also related to the quantitative international trade literature, which has developed tractable models of trade across multiple countries with various dimensions

of heterogeneity (Anderson,1979;Eaton and Kortum,2002;Melitz,2003). These

sem-inal models characterize trade and the distribution of economic activity across countries as a function of exogenous trade costs. A standard prediction of these models is that the relationship between trade flows and costs follows a gravity equation, which has been

documented as one of the strongest empirical regularities in the data (Head and Mayer,

4_{Another related paper is Falvey(1976), who discusses how the transportation sector can} draw away resources from tradables in particular.

2014). We complement this literature by developing a framework in which trade costs are endogenous, in a way that is both tractable and preserves the gravity structure of

trade flows. This relates our paper to Fajgelbaum and Schaal (2020) and Santamaría

(2020), who consider endogenous road construction in multi-location models of

eco-nomic geography, as well asBrancaccio, Kalouptsidi, and Papageorgiou (2020), who

endogenize trade costs in the non-containerized shipping sector. Unlike these papers, we focus on port development as a source of endogenous shipping costs, and solve for the decentralized equilibrium as opposed to the optimal allocation to quantify the effect of port development on trade, the distribution of population, and welfare.

Finally, our paper is related to a large literature studying the effects of transport

in-frastructure improvements.5 In particular, there is a growing empirical literature

study-ing the effects of containerization (Hummels, 2007; Bernhofen, El-Sahli, and Kneller,

2016;Gomtsyan, 2016;Co¸sar and Demir,2018;Holmes and Singer,2018;Altomonte et al., 2018; Brooks et al., 2019) or the role of container shipping networks in world

trade (Wong,2020;Heiland, Moxnes, Ulltveit-Moe, and Zi,2021;Ganapati, Wong, and

Ziv, 2020). Most closely related is Brooks et al. (2019), who study the reduced-form

effects of containerization on local economic outcomes across U.S. counties. Our main contribution to this literature is twofold. First, motivated by the evidence that container-ization dramatically increased land use in ports, this paper highlights the crowding-out effect of containerization and finds sizeable local and global costs stemming from this effect. Second, to the best of our knowledge, this is the first paper seeking to quantify the aggregate effects of port development on global trade and welfare through the lens of a general equilibrium economic geography model.

The paper is structured as follows. In the next section, we describe the main features

of containerized technology. Section 2 discusses the main data sources used in the

analysis. Section 3 presents the reduced form empirical strategy and results, while

Section4introduces the model. Section5revisits the empirics guided by the predictions

of the model. Section6measures the aggregate effects of containerization and Section

7considers alternative explanations for the crowding-out mechanism. In Section8, we

illustrate the effects of targeted port development policies similar to the ‘Maritime Silk

Road’. Finally, Section9concludes.

1

The increased land-intensity of containerization

The introduction of steamships and railroads in the 19th century substantially reduced both water and overland transportation costs. However, transshipment remained slow

and expensive through the middle of the 20th century (Krugman,2011). As a report by

McKinsey highlighted; “The bottleneck in freight transport has always been the

inter-face between transport modes, especially the crucial land/sea interinter-face” (1972, pp. 1-3).

Containerization, that is, the handling of cargo in standardized boxes, was the

break-through innovation that dramatically reduced transshipment times and costs (Hummels,

2007; Rodrigue, 2016). In this section, we show that while substantial transshipment cost reductions were achieved in shipping as a result of containerization, this came at the cost of needing to dedicate much more land to the port.

1.1 The cost – space trade-off in containerization

As late as the mid-1950s, transshipment at seaports was a costly and slow procedure as it entailed handling cargo item-by-item – a process called breakbulk shipping. The reason for this was that cargo came in many different sizes and needed to be handled individually, despite the widespread use of machinery introduced pre-containerziation

(see Panel A of Appendix Figure C.1). The San Francisco Port Commission (1971)

estimated that it took 7 to 10 days to merely discharge cargo from a ship using this

technology. According toBernhofen et al.(2016), two-thirds of a ship’s time was spent

in port. This led to high costs as the capital utilization of ships was low, and the cost of

capital tied up in inventory was high.6

U.S. shippers first started placing cargo into containers in the late 1950s.7

Con-tainerized port technology can be seen in its mature form at the Port of Seattle in 1969

in Panel B of Appendix Figure C.1 (a mere 10 to 15 years after the photos shown in

Panel A were taken). Cargo, packed in standardized containers, is loaded onto and off ships using large, purpose-built cranes situated on the wharf. Large, open areas beside the wharf are used to line up containers.

Containerization substantially reduced transshipment costs for a number of reasons. First, as containers could be handled in a uniform way, loading and unloading times

were vastly reduced. The San Francisco Port Commission (1971) estimated that a

con-tainer ship could be unloaded and loaded in 48 hours or less, a tenth of the previous time spent in port. Similarly, using detailed data on vessel turnaround times for one

anonymized port, Kahveci (1999) estimates that the average time ships spent in port

fell from 8 days to 11 hours as a result of containerization, a reduction of 94%. Second,

the reduction in turnaround time justified investment in much larger vessels (Gilman,

1983). The average size of newly-built container ships increased by 402% between

6_{Industry experts estimated that the handling of cargo at the port accounted for a major share} of freight costs (Levinson,2010). As an example, transshipment costs were estimated to account for 49% of the total transport cost on one route from the U.S. to Europe (Eyre,1964).

7_{Containerized shipping was initially introduced on domestic routes between U.S. ports, but} the technology was rapidly adopted and standardized worldwide in 1967 (Rua,2014).

1960 and 1990.8 Larger ship sizes made it possible to realize even larger cost

reduc-tions through increasing returns to scale in shipping and port handling.Rodrigue(2016,

p. 118) estimates that moving from a 2,500 TEU capacity vessel to one with 5,000 TEU reduced costs per container by 50%.

Adapting ports to containerized technology was not without costs, however. Most importantly, faster turnaround times could only be achieved at the cost of building much

larger terminals. In discussing the ‘challenges’ associated with containerization,

Ro-drigue(2016, p. 118) puts site constraints in the first place, and in particular, the large consumption of terminal space. Containerized terminals need more space as the easy accessibility of containers allows for efficient on- and off-loading. Containers are lined up next to where the ships dock, and space is also needed to rapidly off-load cargo. The increased space requirements of containerized facilities were evident from the start. For example, in a 1971 report, alarm bells were being raised about the inadequacy of San Francisco’s finger piers to accommodate new types of cargo handling; “No pier facili-ties in the Bay Area today are capable of handling the new space requirements on this scale of new and larger container ships. (...) thus more berthing and backup area is

needed” (1971, p. 13). Ports in densely built up areas such as Manhattan and San

Fran-cisco were almost certainly doomed to decline as one observer noted for San FranFran-cisco; “Rows of finger piers adjacent to a densely built up city could not adequately serve container shipping, which involved larger ships that required larger wharves and much

larger areas of open space for loading and unloading” (Corbett,2010, p. 164).

1.2 Evidence for the increased land-intensity of containerization

We present two pieces of quantitative evidence that point to a substantial increase in the land intensity of port technology with the introduction of containerization. First, for the Port of Seattle, we are able to measure the area of the port per volume of traffic

handled around the time of containerization.9 _{Between 1961 and 1973, the port built a}

number of new containerized facilities, increasing the area of the port almost fourfold. Consistent with these investments, by 1973, containerized cargo accounted for 43% of total traffic. While the total volume of traffic handled more than doubled, we calculate that area relative to throughput increased by 90%. This suggests that containerized

8_{These calculations are based on data from the Miramar Ship Index (Haworth,2020). More} details on these data are provided in the Supplementary Material.

9_{More precisely, we use engineering maps for all properties under the ownership of the} Seat-tle Port Authority to calculate the area of the port (excluding the airport and other land not used for seaport activities) and divide by the five-year moving-average of throughput to smooth out year-to-year fluctuations in capacity utilization. A more detailed discussion, including informa-tion on data sources, is included in the Supplementary Material.

technology requires substantially more land per unit of cargo shipped.

Second, we examine the relationship between port area and containerized cargo volume across a large set of cities in our sample. We exploit the availability of high resolution remote sensing data that makes it possible to measure the area of ports in

our sample today.10 _{We match this with data on the cargo composition handled by}

each port using data from Le Journal de la Marine Marchande (JMM) for 2008-09.11

Column (2) in TableB.1shows that, controlling for the total volume of traffic, ports that

handle more containerized cargo are typically larger. In columns (3)-(6), we control for additional potential confounders. The coefficient remains significant and very similar in magnitude when we add controls for the volume of non-liquid and solid bulk handled by the port (e.g., oil, coal or grain, the handling of which may be technologically very different), the GDP per capita of the country (to get at differences in the extent of automation), or even country fixed effects.

To get a better sense of magnitudes, in column (7) we use the share of containerized cargo (while continuing to control for the total volume of cargo). The coefficient of interest is large and statistically significant. Based on this specification, moving from a fully non-containerized port to a fully containerized port is associated with a 75%

increase in port area (i.e., exp(0.5624) − 1), holding the volume of traffic fixed.12

In summary, the historical and quantitative evidence paint a consistent picture. The cost reductions containerization makes possible through faster transshipment times can only be achieved at the expense of dedicating more land per unit shipped to the port.

2

Data

Our analysis builds on a decadal city-level dataset of shipping flows, population, and other economic outcomes for the period 1950-1990. We complement this with GIS data that allows us to calculate geographic characteristics of the city. We review the main variables used in the analysis below and report summary statistics in Appendix Table

B.2. Additional details for all the data used in the paper can be found in the

Supple-mentary Material.

Shipping flows. Crucial to our analysis is a dataset of worldwide bilateral ship

move-ments at the port level for the period 1950-1990 from Ducruet, Cuyala, and Hosni

10_{In particular, for a random set of cities in our dataset, we hand-coded polygons from Google} Earththat we identified as containing port activities. A more detailed description of data and methodology can be found in the Supplementary Material. Unfortunately, the resolution of historical satellite images is not sufficiently high to replicate this exercise for our sample period.

11_{These data were not available for a more recent year.}

12_{The binscatter (plotted in Appendix FigureC.2) visualizes the positive relationship between} the area occupied by the port and the share of containerized cargo.

(2018). An observation is a ship moving from one port to another at a particular point in

time.13 One week samples of these data were extracted from the Lloyd’s Shipping Index,

a unique source that provides a daily list of merchant vessels and their latest inter-port

movements.14 _{We are aware of no previous application in the economics literature.}

These data provide us with rich variation to study the geography of sea-borne trade through the second half of the 20th century. They cover both domestic and interna-tional shipping. Moreover, the data cover a long time period spanning the containeriza-tion revolucontaineriza-tion. We are thus able to compare the effects of port activity on cities both before and after the arrival of the new technology. We know of no other data source that has a similar coverage across time and space, especially at such a detailed level of disaggregation. An important limitation, however, is that we do not observe either the value or the volume of shipment but only bilateral ship movements. From these ship movements, we sum the total number of ships passing through each port, which we call shipping flows.

City population. As we are interested in the economic effects of containerization, we use data on city population worldwide for locations with more than 100,000

inhab-itants from Villes Géopolis (Moriconi-Ebrard, 1994) for each decade between

1950-1990 (Geopolis cities, henceforth). The advantage of these data relative to sources such as the more frequently used UN World Cities dataset is that a consistent and system-atic effort was made to obtain populations for the urban agglomeration of cities (that is, the number of inhabitants living in a city’s contiguous built-up area) as opposed to the administrative boundaries that are often reported in country-specific sources. For example, New York (New York) and Newark (New Jersey) form one ‘city’ according to this definition. We observe population for cities that reached 100,000 inhabitants in any year throughout this period. For most of these cities, we observe population even when the city had fewer than 100,000 inhabitants, leading to potentially important sampling bias. To address this, we will show that our results are robust to using the subset of cities that had already attained 100,000 inhabitants in the first sample year, 1950.

Ports were hand-matched from the shipping data to cities based on whether the port was located within the urban agglomeration of a city in the Geopolis dataset, allowing

for multiple ports to be assigned to one city (Ducruet et al.,2018). We define port cities

in a time invariant manner; a port city with positive shipping flows in at least one year

13_{As such, it is similar to contemporary satellite AIS (Automatic Identification System) data} that tracks the precise movements of vessels around the globe. These type of AIS data are used inHeiland et al.(2021) andBrancaccio et al.(2020).

14_{The data were entered from issues for the first week of May. The data are discussed in more} detail inDucruet et al.(2018).

will be classified as a port city for all years. Of the 2,636 cities in the Geopolis dataset, 553 have at least one port. We label these as port cities. The quantitative estimation covers the full set of 2,636 Geopolis cities (port and non-port cities).

Underwater elevation levels. We use gridded bathymetric data on underwater elevation levels at a detailed spatial resolution (30 arc seconds, or about 1 kilometer at the equa-tor) from the General Bathymetric Chart of the Oceans (GEBCO) to measure sea-depth around the city.

Saiz land rent proxy. While we are not aware of any dataset that covers land rents

glob-ally going back to the 1950s, Saiz (2010) has proposed a geography-based measure

that correlates well with land-rents. This allows us to construct land rent proxies for all cities in our dataset. The ‘Saiz-measure’ is defined as follows: Take a 50 kilometer radius around the centroid of the city. Exclude all sea cells, all internal water bodies and wetland areas, as well as all cells with a gradient above 15%. The remaining cells, as a share of the total cells, can be used as a proxy for land rents. We replicate the

methodology inSaizexactly, using GIS data that have global coverage.

City-level GDP per capita. Data on city-level income levels are needed for the quanti-tative estimation only. We are not aware of readily available sources of GDP per capita data for cities worldwide. For this reason, we estimate GDP per capita for the last year in our sample (1990) for the full sample of 2,636 worldwide cities in the following way. First, we use estimates of city GDP from the Canback Global Income

Distribu-tion Database for a subset of our sample (898 cities) for which data are reported for

1990. We extrapolate GDP per capita for the full sample of cities using the linear fit of the GDP per capita data on nightlight luminosity and country fixed effects, building on a growing body of evidence suggesting that income can be reasonably approximated

using nightlight luminosity data (Donaldson and Storeygard,2016).

3

The reduced form effects of containerization

In this section, we study the local effects of containerization on port cities. To isolate the causal effect, we first develop an exogenous measure of port suitability, and then proceed to discussing our main reduced form findings.

3.1 An exogenous measure of port suitability

Section1discussed the fact that containerization led to larger ship sizes, and that this

in turn required greater depth at the port. Following the previous literature, we think of naturally endowed depth as an exogenous cost-shifter that makes it cheaper for a

port to reach a desired depth through costly dredging (Brooks et al., 2019;Altomonte

naturally endowed depth and depth attained by dredging. Our solution to this relies on using contemporary granular data on underwater elevation levels around the port to isolate the naturally endowed component of depth. In particular, we take all sea cells within buffer rings around the geocode of the port and sum the number of cells that

are ‘very deep,’ which we define as depth greater than 30 feet followingBrooks et al.

(2019). These authors argue that given vessel sizes in the 1950s (pre-containerization),

depth beyond 30 feet conferred no advantage to the port. Below, we will test how reasonable this assumption is by examining pre-trends in shipping.

To operationalize our measure, we need to take a stand on which set of cells around the port to consider. Our aim is to measure depth in areas around the port that are used by ships to navigate and wait for their docking time. We examine the location (using exact geocodes) of stationary ships around the port in a one hour window for

100 random ports in our sample using contemporary data.15 _{The majority of stationary}

ships are located within 5 km, which justifies our baseline measure of port suitability:

the log of the sum of ‘very deep’ cells in a buffer ring 3-5 km around the port.16 _We

examined the effect of depth measured at various buffers and confirmed that the effects are similar in nearby rings, suggesting that the variation we use from the 3-5 km buffer is a representative measure of depth at the port.

Testing for endogenous dredging. The key assumption behind our ability to isolate naturally endowed depth (from depth attained by dredging) is that when ports need to invest in costly dredging, they typically do not dredge entire areas in our buffers, but narrow channels that ships use to navigate to the port. By calculating depth over many sea cells, the vast majority of depth measurements for each port should reflect naturally endowed depth. We test this assumption in the following way. For 100 random ports in our sample, we obtained access to nautical maps from marinetraffic.com which

clearly demarcate the dredged channels that ships use to navigate to the port.17 _We

then constructed a binary variable, ‘Dredging’, that takes the value 1 if a port has a

dredged channel in the 3-5 km buffer ring. Appendix TableB.3 shows the association

15_{These data are from marinetraffic.com and refer to stationary ships near the port captured} between November 4 and 10, 2019, at 12:00-13:00 local time. More details regarding these data and how we choose the buffer around the port are provided in the Supplementary Material. There is a concern that measures of where ships are found around the port today is a poor proxy for where ships were located during our sample period. Partly for this reason, we show that depth measured in the same way at different nearby buffers yields similar results.

16_{There are zeros in the data, that is, there are ports with no cells deeper than 30 feet in the 3-5} km buffer around the port. For this reason, in practice, we use ln(1 +P

i1(depthi ≥ 30f t)), where i denotes a cell.

between this measure and the depth measure. The unconditional association (column (1)) is negative and statistically significant. That is, ports that we measure to be shallow are more likely to have a dredged channel. This is what we would expect to find if our

measure captured naturally endowed depth.18

Balancing checks. We examine the extent to which our measure of exogenous port suitability is correlated with other observables pre-containerization in order to assess the

types of confounders that may bias the results. Appendix TableB.4shows the results. If

greater depth would have led to more shipping even before containerization, we would expect to see a positive coefficient between depth and shipping flows. However, we see that the unconditional measure of depth is negatively correlated with both the level of shipping flows in 1950 (measured in logs), and population in 1950 (measured in logs), indicating that initially small cities had larger depth. In terms of growth rates pre-containerization, depth is weakly positively correlated with population growth between 1950 and 1960 (the coefficient is significant at 10%). This suggests that our depth measure is correlated with small cities that are growing relatively fast, i.e., population convergence. In order to purge our depth measure of this variation, we residualize

it on city population in 1950 (measured in logs).19 _{We re-examine how the part of}

the variation in depth that is uncorrelated with 1950 population, ‘residualized depth,’ correlates with the same observables. Reassuringly, residualized depth is correlated neither with the level of shipping and population in 1950 (the latter by construction), nor with the change in shipping and population between 1950 and 1960. In the empirical analysis, we therefore use the residualized measure of depth as the baseline measure of exogenous port suitability.

Appendix Table B.4 also shows the correlation with other observables.

Residual-ized depth is uncorrelated with all observables we consider, except the Saiz land rent proxy. This is perhaps unsurprising, as arguably similar geographic characteristics de-termine the overland (Saiz measure) and underwater (depth measure) geographic fea-tures around a city. For this reason, we show robustness of all our results to the inclusion of the Saiz land rent proxy interacted with year indicator variables. The final issue con-cerns potential spatial correlation in the depth measure. We will tackle the issue of spatial correlation in the empirics by testing the robustness of our results to using only

18_{Adding continent or coastline fixed effects (columns (2) and (3), respectively) reduces the} size of the negative coefficient and we lose statistical significance in column (3), but the esti-mated coefficients remain negative.

19_{More precisely, we regress the log of depth on the log of population in 1950 and take the} residuals from this regression. Population in 1950 is not observed for 21 out of 553 port cities. For these, we replace 1950 population with the first year in which population is observed, which is generally 1960.

within-region variation, and to adjusting for spatial autocorrelation in the error term by

reporting Conley standard errors (Conley,1999).

3.2 Dealing with land reclamation

Finally, both the depth and the Saiz measure are constructed using contemporary GIS data, which captures natural geography in combination with investments in reclaiming land from the sea. To investigate the extent to which reclamation may introduce

sys-tematic measurement error, we use data fromMartín-Antón, Negro, López-Gutiérrez,

and Esteban(2016) on coastal land reclamation for any purpose.20 _{Appendix TableB.5}

shows that there is somewhat more land reclamation in cities we measure to be more geographically constrained (and hence have higher land rents). This is what we would expect if the Saiz measure was mostly capturing natural geography. The reason for this seems to be that while land reclamation is fairly common (76 out of 553 ports report

some land reclamation), it is typically small relative to the area over which the Saiz

measure is constructed.21 Turning to the association between the depth measure and

the binary indicator of land reclamation, the estimated coefficient is small and never statistically significant.

3.3 Results

Result 1: Depth predicts shipping, but only after 1960. First, we examine whether depth predicts shipping flows during our sample period. We implement this using the follow-ing flexible specification that allows us to examine the timfollow-ing of when depth started to matter for shipping.

ln(Shipit) =

1990

X

j=1960

βj ∗ Depthi∗1(Y ear = j) +

1990

X

j=1960

φj ∗ ln(P opi,1950) ∗1(Y ear = j)

+αi+ δt+ it

The outcome variable of interest, ln(Shipit), is the log of shipping flows observed in

city i at time t. We need to take a stand on the treatment of zeros in the shipping data.22

20_{See the Supplementary Material for a discussion of the data.}

21_{The median size of reclaimed area in the sample for the non-zero observations is 13 square} kilometers, which pales in comparison to the 7850 square kilometers covered in the Saiz mea-sure.

22_{The data contain zeros for two reasons. First, we may observe zeros because of} measure-ment error: small ports with low shipping flows may not register an inter-port movemeasure-ment during the week in which we capture the data. Second, zeros may appear due to the time-invariant definition of port status that we use. We observe zero shipping flows in a particular year if a port was established in the city only after 1950, or if a port shut down in the city during our sample

In the baseline measure, we annualize the weekly counts of ships from the raw data by multiplying the one-week sample of shipping flows we observe by 52. This is primarily so that our results are comparable to regressions we run using model-simulated data

in the quantification exercise in Section 6. Finally, we replace the zeros in the data

with ones and take the natural logarithm of this adjusted annualized count.23 _Depth

i is

the cross-sectional measure of port suitability defined in the previous subsection. We interact this measure with binary indicators for the decades 1960 – 1990 to estimate the time path of how depth affected shipping flows. In addition, we include the full

set of city and year fixed-effects (denoted αi and δt, respectively) as well as the log

of population in 1950 interacted with year indicator variables across all specifications. This is equivalent to using the residualized depth measure in a panel setting. We cluster standard errors at the city level in the baseline to account for the serial correlation of

shocks. We also report Conley standard errors (in curly brackets).24 _{Each β}

j in this

specification estimates the increase in shipping caused by having a deeper port in a given year relative to 1950.

Table1contains the estimated coefficients. Column (1) presents coefficients for the

baseline specification. A number of points should be noted. First, deeper ports did not witness differential growth in shipping flows between 1950 and 1960, consistent with this being a decade in which containerization was just being developed in a few ports around the world. Second, we see an effect of depth in each of the following decades, as containerization was adopted worldwide. The coefficient of interest is much larger and significantly different from zero for the interaction of depth and each year indicator including and after 1970. This is consistent with containerization technology being rolled out in the early 1960s across US ports and worldwide later in the decade, as we

discussed in Section1.

A causal interpretation of the estimated effect of depth relies on the identifying assumption that the time-varying effect of depth is uncorrelated with the error term. The timing of when depth started to matter and the lack of pre-trends provide some evidence that this assumption is plausible. Next, we turn to further testing this result with more demanding specifications. One concern is that many determinants of depth may be

period. Overall, we observe zero shipping flows for 16% of the port-year observations. From examining the data, the zeros seem to be more likely driven by mismeasuring small shipping flows rather than the entry and exit of ports.

23_{In robustness checks discussed below, we show that all of the results presented in this} section are robust to other standard ways of dealing with the zeros. In these, we do not annualize the data in order to verify that this transformation does not drive the results.

24_{As these are typically very close to the clustered standard errors, we only report them for} the main results for easier readability of the tables. We allow for spatial correlation at distances up to 1,000 km and set the spatial decay function to be linear.

Table 1: Depth predicts shipping flows, but only after 1960 Dependent variable: ln(Shipment)

Independent variables (1) (2) (3) (4) (5) Depth × post 1970 0.247*** (0.059) {0.052} Depth × 1960 -0.051 0.029 0.050 -0.055 (0.063) (0.069) (0.066) (0.068) Depth × 1970 0.222*** 0.233*** 0.278*** 0.213*** (0.069) (0.077) (0.082) (0.071) Depth × 1980 0.188** 0.212** 0.291*** 0.192** (0.079) (0.085) (0.090) (0.081) Depth × 1990 0.255*** 0.222** 0.312*** 0.283*** (0.086) (0.087) (0.099) (0.087) Observations 2765 2765 2765 2360 2765 R-squared 0.126 0.248 0.131 0.142 0.126 Number of cities 553 553 553 472 553 Year FE 3 3 3 3 3 City FE 3 3 3 3 3 Population 1950 × Year 3 3 3 3 3 Coastline × Year FE _{5 3 5 5 5} Saiz × Year _{5 5 3 5 5} GDP pc (country) × Year _{5 5 5 3 5}

Notes:‘Depth’ indicates the port suitability measure. It is interacted with decade dummies or an indicator variable for decades including and after 1970, as indicated. Standard errors clustered at the city level in parentheses, Conley standard errors to adjust for spatial correlation in curly brackets. *** p<0.01, ** p<0.05, * p<0.1 (significance refers to clustered standard errors).

spatially correlated and if true, the estimates could be hard to disentangle from broader regional trends. To this end, column (2) adds the full set of ‘coastline’ by year-fixed effects to examine the extent to which our identifying variation relies on cross-regional

variation.25 _{Note that this set of fixed effects subsumes continent by year fixed effects.}

Column (3) adds the Saiz land rent proxy interacted with year indicators to capture trends driven by the time-varying effect of land rents. Column (4) adds country GDP per capita (measured in 1960) interacted with year indicators to control for potentially

differential growth trends across initially rich and poor countries.26 _{The coefficients are}

remarkably stable.

Based on these results, we introduce a ‘containerization’ treatment indicator that turns on in years including and after 1970. This yields a single coefficient that estimates the differential effect of depth on shipping after the onset of containerization. Column (5) shows the results. Cities endowed with more depth, and hence more suitable to containerized technologies witnessed disproportionate increases in their shipping flows

after containerization. Panel A in Appendix FigureC.3shows the coefficient of interest

remains fairly stable as we drop continents one at a time, underscoring that no single re-gion appears to be driving the results. The coefficient becomes somewhat smaller when we drop North America, which is in line with the United States being the birthplace and an early adopter of containerization.

Panel A in Appendix Table B.6 contains further robustness checks. First, we test

robustness to different data construction choices. In particular, we examine different ways of treating zero shipping values, different ways of defining the depth measure for the handful of ports that are located far inland from the coastline and restricting the sample to the subset of cities that had already attained 100,000 inhabitants by 1950 to examine sample selection bias. The coefficient of interest remains similar in magnitude and highly significant across all these checks. We now turn to examining how this boom in shipping affected city population.

Result 2: The local causal effect of shipping on population is not distinguishable from

zero. We estimate the effect of shipping on population using the following specification;

ln(P opit) = β ∗ ln(Shipit) +

1990

X

j=1960

φj∗ ln(P opi,1950) ∗1(Y ear = j)+αi+ δt+ it (1)

where ln(P opit) is the natural logarithm of population in city i at time t, and all other

variables are as previously defined. The identification challenge is that the shipping flows of a city are endogenous. Our main worry is reverse causality: fast growing cities

‘Pacific Ocean’) or body of water (e.g., ‘Great Lakes’) and further disaggregate oceans by conti-nent. This yields 22 coastlines worldwide. Examples are ‘Mediterranean – Europe’ and ‘North America – Atlantic’.

26_{We use the 1960 (pre-containerization) measure of country GDP per capita as this is} ob-served for a larger set of countries than for 1950.

will also witness increases in their shipping flows. Our solution is to isolate variation in shipping using the binary version of our containerization treatment defined in the previ-ous section: we interact the cross-sectional measure of depth with an indicator variable that takes the value of one in years including and after 1970. We cluster standard errors at the city level and we also report Conley standard errors for our main results.

Table 2: The local causal effect of shipping on population is not distinguishable from zero

Panel regression Long difference

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Indep. Variables ln(Pop) ln(Pop) ln(Ship) ln(Pop) ln(Ship) ln(Pop) ∆ln(Pop) ∆ln(Pop) ∆ ln(Ship) ∆ln(Pop) ln(Shipment) 0.013*** 0.015 0.030*** 0.035 (0.005) (0.049) {0.005} {0.039} ∆ ln(Shipment) 0.013 0.006 0.052 0.022 (0.009) (0.073) {0.014} {0.115} Depth 0.272*** 0.002 0.134*** 0.003 (0.086) (0.020) Depth × post 1970 0.268*** 0.004 0.143*** 0.005 (0.058) (0.013) Depth × 1960 -0.042 -0.003 (0.064) (0.008) Depth × 1970 0.246*** 0.007 (0.069) (0.013) Depth × 1980 0.213*** -0.002 (0.079) (0.017) Depth × 1990 0.280*** 0.002 (0.086) (0.020) Observations 2734 2734 2734 2734 2734 2734 531 531 531 531 Number of cities 552 552 552 552 552 552 Year FE 3 3 3 3 3 3 5 5 5 5 City FE 3 3 3 3 3 3 5 5 5 5 Population 1950 × Year 3 3 3 3 3 3 5 5 5 5 Population 1950 5 5 5 5 5 5 3 3 3 3

Specification OLS 2SLS FS RF dyn FS dyn RF OLS 2SLS FS RF

KP F-stat 21.13 9.98

Notes: ‘Depth’ indicates the port suitability measure. It is interacted with decade dummies or indicator variables for decades including and after 1970, as indicated. Standardized coefficients in italics underneath the baseline coefficients. Notation for specification as follows: ‘FS’ refers to the first stage, ‘RF’ to the reduced form, ‘dyn FS’ to the fully flexible first stage and ‘dyn RF’ to the fully flexible reduced form. Standard errors clustered at the city level in parentheses, Conley standard errors to adjust for spatial correlation in curly brackets. *** p<0.01, ** p<0.05, * p<0.1 (significance refers to clustered standard errors).

the effects of interest using all years in the sample, while columns (7) to (10) show the estimates from the long-differenced specification. Turning first to the full panel

specification, the OLS specification of equation (1) shows that the association between

shipping and population is small, positive and statistically different from zero (coef-ficient 0.013, se. 0.005). The 2SLS estimate in column (2) shows a similarly sized coefficient but we cannot reject zero (coefficient 0.015, se. 0.049). To assess magni-tudes, we report the standardized ‘beta’ coefficients for our effects of interest in italics underneath the estimated regression coefficients. These make clear that while the OLS may be statistically significant, the magnitudes of both the OLS and the 2SLS estimates are economically negligible. A one standard deviation increase in shipping leads to a 0.03 (OLS) or 0.035 (2SLS) standard deviation increase in population. Columns (3) and (4) show the first stage and reduced form respectively. These make clear why the results are indistinguishable from zero. While the first stage is strong (the Kleibergen-Paap F-statistic is 21.13), there is no reduced form relationship between depth and population

(the reduced form coefficient is 0.004, se 0.013).27 _{Column (6) shows the full time path}

of effects for the reduced form. These make clear that the statistically insignificant co-efficient in the 2SLS estimate does not stem from the fact that population is sluggish to adjust. The time path of the coefficients shows no discernible trend, and there is no clear difference in population growth post-containerization for deeper ports. All of the coefficients are estimated to be very close to zero (the one ‘furthest’ away from zero is 0.007), the coefficients are never close to statistical significance and in two of the five decades, the estimated reduced form coefficient is negative, suggesting that if anything, deeper ports were growing at a slower rate than shallower ones in some decades.

While the large standard errors typical of 2SLS estimation make a definitive an-swer difficult, there are several reasons why we believe that the most reasonable in-terpretation of our results is that shipping booms caused by containerization led to no discernible effects on population. First, the standardized ‘beta’ coefficients make clear that the magnitudes of both the OLS and the 2SLS estimates are economically negligi-ble. Second, if we examine the long-differenced specification in columns (7) to (10), neither the OLS nor the 2SLS estimate is significantly different from zero, and both standardized beta coefficients again show an economically negligible effect. In fact, the 2SLS coefficient estimate is smaller – it is less than half the size estimated in the full panel, consistent with the fact that it was population observations from earlier years that drove the point estimate in the full panel specification. While the long-differenced specification has the disadvantage of using fewer observations, it has the advantage that

27_{The specification here is identical to that in Table1, but the sample size shrinks slightly as} we lose those observations where population is unobserved in some years (1% of the sample).

it examines the long-run effects of the shipping boom on population, once the latter has had time to adjust.

Third, we subject the 2SLS specification to the same set of robustness checks con-ducted above: the inclusion of coastline by year fixed effects, controlling for the time varying-effect of the the Saiz land-rent proxy and GDP per capita (Appendix Table

B.7). Despite the demanding nature of these specifications, the first stage remains

suf-ficiently strong (the Kleibergen-Paap F-statistic is always above 10) and the estimated 2SLS coefficient is never statistically different from zero. In fact, in two out of three cases, the estimated coefficient is negative. Fourth, no single continent drives this result

(Appendix FigureC.3, panel B, plots the estimated coefficient dropping continents one

at a time). Appendix Table B.6, panel B, shows that the results are robust to the same

set of additional robustness checks to data choices performed for Result 1.

We view the null effect on population as a surprising finding. Intuition and standard

models (Co¸sar and Fajgelbaum, 2016; Nagy, 2020b;Fajgelbaum and Redding, 2018)

would both suggest that a boom in shipping should make a location more attractive for households and firms, as they can access consumers and producers more cheaply (the ‘market access effect’), leading to an inflow of population. Indeed, the past literature has found that these types of positive shocks to a city’s accessibility tend to lead to a

boom in local population (Bleakley and Lin(2012);Brooks et al.(2019);Campante and

Yanagizawa-Drott(2018).28 While comparing the economic size of the effect in these papers relative to ours is difficult given the different contexts and different ‘treatments,’ these papers all show that their effect is economically meaningful, while making the same claim with our results would be difficult.

What can explain the difference between our findings and previous work? One no-table difference in our setting is the increased land intensity of port activities induced

by containerization discussed in Section1. In the last part of this section, we examine

the extent to which we can detect the effects of this in our data. We also note that in

Section6, we will use our quantified model to return to the question of what magnitude

one would expect in our setting in the absence of the crowding out mechanism.

Result 3: Containerization increased shipping more in low rent cities.To the extent that

28_{The paper closest to our setting isBrooks et al.(2019), who study the effect of} container-ization on the population of U.S. counties located nearby. They find a positive and statistically significant effect of containerization on local population. Though the two settings are difficult to compare as we study cities around the world, we think one crucial difference is that while we examine the effects on cities, their unit of analysis is a county. As we argue below, the most likely mechanism driving the null result is that the land intensity of port technology acts as an important opposing force crowding out population. This mechanism is more likely to be detectable at the generally finer level of spatial resolution that we examine.

land prices affect where port development takes place, we would expect low land-rent cities to be more attractive places for containerized ports all else equal, as the opportu-nity cost of port development in these locations is low. We test for this by examining the heterogeneity of the depth-shipping relationship from Result 1 using the following specification;

ln(Shipit) = β ∗ Depthi∗1(Y ear ≥ 1970) + γ ∗ Depthi∗ Renti∗1(Y ear ≥ 1970)

+η ∗ Renti∗1(Y ear ≥ 1970) +

1990

X

j=1960

φj∗ ln(P opi,1950) ∗1(Y ear = j)

+αi+ δt+ it (2)

where Rentiis the Saiz land rent proxy for city i, and all other variables are as defined

above.29 _{The coefficient of interest is γ – that is, we are interested in the interaction}

between our depth measure and the Saiz land rent proxy (interacted with the ‘container-ization’ treatment variable that turns on in 1970). We have defined the Saiz measure such that higher values correspond to less area that can be developed, implying high land rents. Note that this is a fully saturated specification in that we allow both depth and the Saiz measure to have their own time trend break in 1970. We plot the marginal

effect of depth at different values of the Saiz measure in Figure 1a (the

correspond-ing estimates are presented in Appendix TableB.8). Consistent with the land intensive

nature of containerized technology, the coefficient of interest, γ, is negative, large and statistically different from zero (coefficient -0.707, se. 0.323). Cities with exogenously deeper ports witnessed increased shipping flows after 1970, but disproportionately more so in low land rent cities.

Panel C in Appendix FigureC.3explores the heterogeneity of the result by dropping

continents one at a time. The effect is consistently negative. We perform the same set of

robustness checks for this result as for previous ones (see Appendix TablesB.6, panel

C, andB.8). The results are largely robust to these specifications, as our coefficient of

interest, γ, remains negative and economically large throughout all these checks, though in two especially demanding specifications the level of significance drops slightly below 10%.

In Appendix Table B.9 we provide additional evidence that land prices matter for

where port activity takes place by examining the exact location of ports within cities.30

29_{We report standard errors clustered at the city level, as well as Conley standard errors in} curly brackets.

(a) Data (b) Model-simulated data

Figure 1: Containerization increased shipping more in low land-rent cities

Notes: Panel A shows the estimated γ coefficient from equation (2) evaluated at different values of the Saiz land rent proxy. Panel B shows the same estimated coefficient using model-simulated data evaluated at different values of the counterfactual land rents.

We show that during this time period, ports moved further from the centroid of the city

towards the outskirts, where land prices are typically lower (Duranton and Puga,2019).

This is particularly striking for the subset of cities in which a new port was built (e.g., in Sydney, Australia). In these cases, the new port was located on average 9 km further from the centroid of the city than the old port.

The land rent heterogeneity result suggests that the land-intensive nature of con-tainerized technology is an empirically important determinant of where concon-tainerized port infrastructure was developed. Armed with this evidence, we now turn to writing down a quantitative spatial model that captures many realistic features of port infras-tructure development, including, but not limited to, the land price mechanism. We note that of course there are mechanisms other than the land-intensity of containerization that could account for the crowd-out of population. We explore three prominent ones in

Section7, after presenting both the empirics and the quantified model.

4

A model of cities and endogenous port development

To measure the aggregate effects of port development, we develop a rich and flexible quantitative general equilibrium model of trade across cities that captures both the ben-efits and costs of port development.

4.1 Setup

The world consists of S > 0 cities, indexed by r or s. An exogenously given subset of cities are port cities, while the rest are non-port cities. We make the Armington assumption that each city produces one variety of a differentiated final good that we

1953 and 2017. We provide details on the data used for this exercise, and in particular, on how we calculate city centroids in the Supplementary Material.

also index by r or s (Anderson, 1979). Each city belongs to one country, and each country is inhabited by an exogenous mass of workers who choose the city in which they want to live. Mobility across cities is, however, subject to frictions.

4.1.1 Workers

Each worker owns one unit of labor that she supplies in her city of residence. The utility of a worker j who chooses to live in city r is given by

uj(r) = " _S X s=1 qj(r, s) σ−1 σ #_σ−1σ a (r) bj(r) (3)

where qj(r, s) is the worker’s consumption of the good made in city s, a (r) is the level

of amenities in city r, and bj(r) is an idiosyncratic city taste shifter. σ > 1 is the

elasticity of substitution across goods.

The dispersion of bj(r) represents the severity of cross-city mobility frictions that

workers face, similar to Kennan and Walker (2011) and Monte, Redding, and

Rossi-Hansberg (2018). For tractability, we assume that bj(r) is drawn from a Fréchet

dis-tribution with shape parameter 1/η and a scale parameter normalized to one. Hence, a larger value of η corresponds to more severe frictions to mobility.

4.1.2 Landlords

Each city r is also inhabited by a positive mass of immobile landlords who own the exogenously given stock of land available in the city. We normalize the stock of land

available in each city to one.31 Landlords have the same preferences over goods as

workers. They do not work but finance their consumption from the revenues they collect after their stock of land.

Each landlord is small relative to the total mass of landlords in the city and hence thinks that she cannot influence prices. Yet the mass of landlords is small enough that the population of each city can be approximated well with the mass of workers who choose to reside in the city.

In non-port cities, landlords rent out their land to firms that produce the city-specific good. In port cities, landlords can also use part of their land to provide transshipment services. The more land they use for transshipment services, the more the cost of

trans-31_{We could allow the stock of available land to vary across cities. This more general setup} is isomorphic to our current model, except that, instead of productivity in the city-specific good sector, a combination of the stock of land and productivity enters the model’s equilibrium con-ditions. In other words, the city productivity levels we identify from our current model reflect not only productivity per se, but also the stock of available land. This fact, however, does not affect our quantitative results as we keep productivity levels fixed in our model simulations.

shipping a unit of a good decreases. The landlord can charge a price for the transship-ment service she provides. Competition among port city landlords drives down this price to marginal cost. Hence, profits from transshipment services are zero in

equilib-rium.32

4.1.3 Production

Firms can freely enter the production of the city-specific good. Hence, they take all prices as given and make zero profits. Production requires labor and land. The repre-sentative firm operating in city r faces the production function

q (r) = ˜A (r) n (r)γ(1 − F (r))1−γ

where q (r) denotes the firm’s output, ˜A (r) is total factor productivity in the city, n (r)

is the amount of labor employed by the firm, and F (r) is the share of land that landlords in the city use for transshipment services (thus, F (r) = 0 in non-port cities). Hence, 1 − F (r) is the remainder of land that landlords rent out to firms for production, and γ and 1 − γ correspond to the expenditure shares on labor and land, respectively.

We incorporate agglomeration economies by allowing total factor productivity to depend on the population of the city, N (r):

˜

A (r) = A (r) N (r)α

where A (r) is the exogenous fundamental productivity of the city, and α ∈ [0, 1 − γ] is

a parameter that captures the strength of agglomeration economies.33 _{The representative}

firm does not internalize the effect that its employment decision has on local population. Hence, it takes N (r) as given.

4.1.4 Shipping and port development

Firms in city r can ship their product to any destination s ∈ S. Shipping is, however, subject to iceberg costs: if a firm i from city r wants to ship its product over a route ρ that connects r with s, then it needs to ship T (ρ, i) units of the product such that one

unit arrives at s. Shipping costs consist of a component common across firms ¯T (ρ), as

32_{In Section 6, we show that the aggregate gains from containerization remain similar in an} alternative framework in which landlords have market power and thus can make profits. We provide a detailed description of this alternative framework in the Supplementary Material.

33_{We make the assumption α ≤ 1 − γ to guarantee that agglomeration forces are not} over-whelmingly strong in the model. Estimates of the land share, 1 − γ, tend to be substantially above estimates of agglomeration externalities α. In particular, our calibration involves set-ting α to 0.06 (a standard value used in the literature) and 1 − γ to 0.16 based onDesmet and Rappaport(2017).

well as a firm-specific idiosyncratic component  (ρ, i) that is distributed i.i.d. across

firms and shipping routes:34

T (ρ, i) = ¯T (ρ)  (ρ, i)

For tractability, we assume that  (ρ, i) is drawn from a Weibull distribution with shape parameter θ and a scale parameter normalized to one. Firms only learn the realizations of their idiosyncratic cost shifters after making their production decisions. Therefore, they make these decisions based on the expected value of shipping costs,

E [T (ρ, i)] = ¯T (ρ) E [ (ρ, i)] = ¯T (ρ) Γ θ + 1

θ 

.

After learning  (ρ, i), they choose the route that minimizes their total shipping costs. Certain shipping routes involve land shipping only (land-only), while others involve a combination of land and sea shipping through a set of ports (land-and-sea). Land-only shipping is Land-only available between cities that are directly connected by land. The common cost of land-only shipping between cities r and s is an increasing function of the minimum overland distance between the two cities, d (r, s):

¯

T (ρ) = 1 + φς(d (r, s))

The cost of land-and-sea shipping depends on the set of ports en route. In particular,

the common cost of shipping from r to s through port cities p0, ..., pM takes the form

¯ T (ρ) = [1 + φς(d (r, p0))] [1 + φς(d (pM, s))] M −1 Y m=0 [1 + φτ(d (pm, pm+1))] M Y m=0 [1 + O (pm)]

where φς(d (r, p0)) corresponds to the overland shipping cost between the origin and

the first port en route p0, and φς(d (pM, s)) corresponds to the overland shipping cost

between the last port en route pM and the destination. φτ(d (pm, pm+1)) denotes the

sea shipping cost between ports pm and pm+1, a function of the minimum sea distance

between the two ports, d (pm, pm+1). Finally, O (pm) denotes the price that the firm

needs to pay for transshipment services in port city pm.35

34_{The assumption of idiosyncratic shipping cost shifters followsAllen and Atkin(2016) and}

Allen and Arkolakis(2019), and allows us to tractably characterize shipping flows with a large number of cities. In the alternative case with no idiosyncratic shifters, applied in Allen and Arkolakis (2014) andNagy (2020a), finding optimal shipping flows is computationally more demanding.